Testing for covariance stationarity in stock market data

Abstract

This paper proposes several non-parametric tests for covariance stationarity and applies them to common stock return data from 1834–1987. Recursive variance plots, post-sample prediction tests, Cumulative Sum (henceforth, CUSUM) tests and modified scaled range tests all show strong non-stationarity in stock returns, primarily due to the large increase in volatility during the Great Depression. These tests should be useful as diagnostics for data where the assumptions underlying the desired statistical procedure require stationarity.